Facial Emotion Detection Project

Overview

This project focuses on building a deep learning model to detect facial emotions from images. The model categorizes emotions into four classes: happy, sad, neutral, and surprise. The project leverages transfer learning using the VGG16 model and hyperparameter tuning to achieve optimal performance.

Problem Definition

Facial emotions constitute a significant part of human communication. Understanding these emotions can enhance human-computer interaction, improve health diagnostics, and enable intuitive AI systems. The goal of this project is to train a model that accurately classifies human emotions based on facial expressions.

Dataset

The dataset consists of three folders: train, validation, and test. Each folder contains images categorized into four subfolders:

• happy: Images of people with happy expressions.
• sad: Images of people with sad expressions.
• neutral: Images of people with neutral expressions.
• surprise: Images of people with surprised expressions.

Methodology

1. Data Preprocessing:

   • Images are resized to 48x48 pixels and normalized.
   • Data augmentation techniques like random flipping and rotation are applied.

2. Model Architecture:

   • The project uses the VGG16 model pre-trained on ImageNet as the base model.
   • Additional layers are added for fine-tuning, including dense layers, batch normalization, and dropout layers.

3. Hyperparameter Tuning:

   • The keras_tuner library is used to optimize the number of neurons in dense layers.
   • The Hyperband algorithm is employed to find the best configuration.

4. Training:

   • The model is trained using the Adam optimizer with a learning rate scheduler (ReduceLROnPlateau).
   • Early stopping is used to prevent overfitting.

5. Evaluation:

   • The model's performance is evaluated on the test set using accuracy, confusion matrix, and classification reports.

Results

• The final model achieved a test accuracy of approximately 72%.
• The confusion matrix and classification report highlight the model's performance across all emotion classes.

Visualizations

• Training and validation accuracy/loss plots are generated to monitor the model's performance over epochs.
• A heatmap of the confusion matrix is plotted to visualize the classification results.

How to Run

1. Clone the repository and navigate to the project directory.
2. Ensure the dataset is downloaded and placed in the Facial_emotion_images folder.
3. Run the FINAL_SUBMISSION.ipynb notebook to train and evaluate the model.

Dependencies

• Python 3.10
• TensorFlow 2.x
• Keras
• Keras Tuner
• Matplotlib
• Seaborn
• NumPy
• PIL

Acknowledgments

• The VGG16 model is pre-trained on the ImageNet dataset.
• The dataset used for this project was provided via Olympus.

Future Work

• Extend the model to detect additional emotions.
• Experiment with other transfer learning models like ResNet or EfficientNet.
• Deploy the model as a web application for real-time emotion detection.
• Optimize the model for faster inference on edge devices.
• Update tests to use F1-score